Abstract: To address challenges such as significant scale variations, high aspect ratios, dense arrangements, and complex backgrounds in ship target detection from remote sensing images, this paper proposes an improved YOLOv7-based algorithm. Using YOLOv7 as the baseline network, the prior anchor generation algorithm is optimized for the dataset. A long-edge representation method combined with circular smooth labeling is introduced to detect ship targets with uncertain rotation directions. The YOLOv7 network is enhanced by embedding both the GAM（Global Attention Mechanism） and Sim AM（Simple Attention Mechanism） modules, which effectively suppress interference from complex background regions in remote sensing images and improve target detection accuracy. Additionally, the coordinate loss function is optimized to accelerate model convergence. Experimental results on the DOTA-ship and HRSC2016 datasets for both single-class and multi-class detection tasks show m AP values of 86. 1%, 97. 7%, and 87. 1%, respectively-representing improvements of 7. 8%, 4. 6%, and 7. 9% over the original YOLOv7 model. These results validate the effectiveness and superiority of the proposed method.